Spinopelvic X-ray demo — OpenSpineConsortium

Reference case —

Lateral radiograph of the lumbar spine and pelvis

Your radiograph drop, paste or choose a file

Drop, paste or choose a lateral film

Ctrl-V / ⌘V pastes an image straight from the clipboard — a screenshot of a PACS window works. Drag a file anywhere onto this panel, or

uses your camera — nothing leaves the page

The model runs in this tab. The image is never uploaded, and this page has no server that could receive one.

model not loaded

Spinopelvic measurements

Reference case — click a parameter to toggle it.

Landmarks

Four corners per vertebra, in the order the model predicts them. Hover a row to find it on the film; hover a marker for its name.

Model

Weights YOLO11m-Pose · 1024

Levels

Levels are an assumption about the film, not a model output. Pick a region to name them; angles that need named endplates appear once you do.

Framing

Confidence 0.50

Trained on 1398 lateral lumbar films (BUU-LSPINE) at fixed hyperparameters with every augmentation zeroed. Corner order is sup_ant, sup_post, inf_ant, inf_post and it is handed — mirror the film if anterior and posterior come out swapped.

On a long or standing film the lumbar spine is a third of the frame, and a single pass at 640 px finds nothing: measured across 40 re-framed test films, detections fall from 6.0/6 to 0 once the spine drops below about two-thirds of the height. Tiling the film into overlapping square windows holds 5.4–5.7 of 6 at every framing down to 0.20, with corner error flat at 0.28–0.33 % of the diagonal. Auto takes the single pass and only pays for tiles when it comes up short.

Finding the femoral head centre

PI and PT are undefined without the bicoxofemoral axis — the midpoint of the two femoral head centres (Legaye & Duval-Beaupère, Eur Spine J 1998). The head is very nearly a sphere, so its projection is a circle and the landmark is the centre of that circle: trace the subchondral cortical arc and mark its centre of curvature, not the brightest shadow — overlap with the acetabulum puts the densest region medial to the true centre.

Drawn on the DRR of this case, so the marked centre is ostk's own 3-D sphere fit projected rather than anyone's opinion. A DRR integrates along the bicoxofemoral axis, so its two heads superimpose exactly — the second head in panel C is therefore a schematic, at the same radius and offset front-to-back, which is how rotation separates them on a real film.

What the film has to contain

SS and LL need only the L1 and S1 endplates, so a coned lumbar lateral is enough and both are reported. PI and PT additionally need the femoral head centres; the model has no landmark for them that survives on a real film, so they are reported as n/a rather than estimated.

The reference pane is a DRR — the CT from the other tab integrated along the patient's true bicoxofemoral axis. Its landmarks were fitted in 3-D and projected, which is why its sacral plate is the true endplate and not the silhouette of the sacral ala.

Research & educational use only. A research demonstration — not a medical device and not for clinical or diagnostic use. Measurements and simulated post-operative results are illustrative and must not guide patient care. Sample imaging is de-identified research data.